Strong-robustness SLAM method and system based on improved visual feature extraction and matching optimization
By improving the SLAM method of visual feature extraction and matching optimization, the problem of the traditional SLAM method degradation in dynamic environments is solved, and higher accuracy, robustness and adaptability are achieved, and suitable for complex scenarios.
Patent Information
- Application Number
- CN202411726677.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-28
- Publication Date
- 2025-03-14
AI Technical Summary
In the presence of dynamic objects, traditional SLAM methods can easily lead to map drift, incorrect matching and instability, resulting in reduced accuracy.
A strong robust SLAM method based on improved visual feature extraction and matching optimization is adopted, and the SLAM model is constructed through the SuperPoint algorithm, including multi-scale feature fusion module, attention module, triple loss module, random consistent algorithm module and comprehensive feature fusion module, to handle the fusion and optimization of multi-feature and multi-source data in complex environments.
It significantly improves the accuracy, robustness and adaptability of visual inertial SLAM, reduces the impact of external interference on the system, and improves the performance of the system in dynamic and unstructured environments.
Smart Images
Figure CN119625333A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of simultaneous positioning and map construction, in particular to a highly robust SLAM method and system based on improved visual feature extraction and matching optimization. Background Art
[0002] SLAM (Simultaneous Localization and Mapping) is an important research field that aims to enable applications such as mobile robots, self-driving vehicles, drones, and augmented reality to achieve autonomous navigation and map construction in unknown environments. Among them, the visual SLAM (vSLAM) system is an important field of SLAM system research because it does not rely on external infrastructure, is suitable for multi-scenario platforms, does not require prior information, and is real-time.
[0003] Traditional SLAM methods usually assume that the environment is static, that is, the objects in the map do not change position or state during the SLAM process. This is reasonable for most mobile robot applications, but in the presence of dynamic objects, dynamic objects will cause map drift, incorrect matching and instability, resulting in serious accuracy degradation in traditional SLAM methods. Therefore, it is crucial to study a method to remove moving objects during SLAM operation in real scenarios. Summary of the invention
[0004] In order to address the deficiencies in the prior art, the present invention provides a highly robust SLAM method and system based on improved visual feature extraction and matching optimization, which can handle the fusion and optimization of multi-feature and multi-source data in complex environments, is suitable for complex indoor and outdoor scenes, significantly improves the accuracy, robustness and adaptability of visual inertial SLAM, and reduces the impact of external interference on the system.
[0005] In order to achieve the above object, the specific scheme adopted by the present invention is: a strong robust SLAM method based on improved visual feature extraction and matching optimization, comprising the following steps:
[0006] Collect environmental images in real time and input them into a pre-trained SLAM (Simultaneous Localization and Mapping) model. The SLAM model is built based on the SuperPoint algorithm. The SLAM model includes a multi-scale feature fusion module, an attention module, a triplet loss module, a random consensus algorithm module, and a comprehensive feature fusion module. The feature encoder of the SuperPoint algorithm is used to extract features from the environment image to obtain multiple original feature maps of different depths; A multi-scale feature fusion module is used to generate multiple hierarchical feature maps of different depths from the original feature map, and a feature connection relationship is established between the hierarchical feature maps to obtain fused features; The attention module is used to adaptively weight the hierarchical feature map to obtain an enhanced feature map; The feature point decoder of the SuperPoint algorithm is used to detect feature points from the enhanced feature map, and the descriptor decoder of the SuperPoint algorithm is used to generate feature descriptors corresponding to the feature points; The triplet loss module is used to optimize feature points and feature descriptors based on distance; The random consensus algorithm module is used to reorder all feature points according to the confidence of the optimized feature points, and multiple feature points are selected as backup data points; Generate a dynamic environment model based on all the alternate data points; The dynamic environment model is optimized based on the preset set constraints using the comprehensive feature fusion module.
[0007] As a further optimization of the above-mentioned strong robust SLAM method based on improved visual feature extraction and matching optimization: a specific method of using a multi-scale feature fusion module to generate multiple hierarchical feature maps of different depths from the original feature map and establish a feature connection relationship between the hierarchical feature maps includes: A feature pyramid network is used to construct a feature map pyramid based on the original feature map. The feature map pyramid includes multiple hierarchical feature maps. The hierarchical feature maps are upsampled and laterally connected to the original feature maps: P i =f(C i )+UpSample(P i+1 ); Among them, P i is the hierarchical feature map of level i, C i is with P i The corresponding original feature map, UpSample(P i+1 ) is upsampling, and f(·) is the convolution operation; Constructing fusion features
[0008] As a further optimization of the above-mentioned strong robust SLAM method based on improved visual feature extraction and matching optimization: the method of adaptively weighting the hierarchical feature map using the attention module includes: Perform global average pooling on each channel in the hierarchical feature map to obtain a global description of each channel: Among them, F∈R H×W×Cis the input hierarchical feature map, H, W, C represent the height, width and number of channels of the feature map respectively, z c Represents the global description of the c-th channel; The result of global pooling is passed through two fully connected layers in sequence. The first fully connected layer is used to reduce the number of channels of the hierarchical feature map, and the second fully connected layer is used to increase the number of channels after the reduction: s = σ(W2δ(W1z)); Where z∈R C is the global information of all channels, and is the weight matrix of the fully connected layer, δ(·) is the ReLU activation function, σ(·) is the Sigmoid activation function, ensuring that the output weight value is in the range of (0, 1), s∈R C is the adaptive weight of each channel; Adaptively weight hierarchical feature maps: where s c is the adaptive weight of the cth channel, To enhance the feature map.
[0009] As a further optimization of the above-mentioned strong robust SLAM method based on improved visual feature extraction and matching optimization: there are three types of feature descriptors, namely Anchor, Positive and Negative. After the feature descriptors are generated, the feature descriptors are converted into feature vectors, namely f A ,f P ,f N .
[0010] As a further optimization of the above-mentioned robust SLAM method based on improved visual feature extraction and matching optimization: the triplet loss function of the triplet loss module is: L(A,P,N)=max{0,d(A,P)-d(A,N)+α}; d(A,P) is the distance between Anchor and Positive, d(A,N) is the distance between Anchor and Negative, and α is the predefined safety margin; With f A Based on the distance metric function, the similarity between feature descriptors is calculated: Among them, P·P2 is the Euclidean norm; The average value of all triple combinations is obtained by minimizing the triple loss function: Where i represents the triple combination that traverses all training samples.
[0011] As a further optimization of the above-mentioned strong robust SLAM method based on improved visual feature extraction and matching optimization: the method to minimize the triplet loss function is the gradient descent method, specifically: in, and They are the gradients of Anchor, Positive and Negative respectively.
[0012] As a further optimization of the above-mentioned strong robust SLAM method based on improved visual feature extraction and matching optimization: a method for reordering all feature points according to the confidence of the optimized feature points using a random consensus algorithm module includes: calculating the confidence of all feature points; Sort all feature points in order from high to low confidence and form an ordered set of feature points: D sorted ={d1,d2,...,d n}, so that p1≥p2≥…≥p n ; Among them, p i is the confidence value of the i-th feature point, and p i The value range of is [0, 1]; The first m feature points are selected from the ordered set of data points as backup data points.
[0013] As a further optimization of the above-mentioned strong robust SLAM method based on improved visual feature extraction and matching optimization: the method of generating a dynamic environment model based on all spare data points includes: All candidate data points are used to fit the basic model. The fitting method is: Among them, θ is the model parameter, x i ,y i are the input and output numbers respectively, f(x i ; θ) represents the use of model parameters θ for input x i Function for making predictions, argmin θ It means finding the parameter set θ that minimizes the error; Calculate the error between all feature points and the basic model, and screen the feature points based on the error to select the internal points relative to the basic model; If the number of filtered inliers reaches a preset threshold, the basic model is used as the dynamic environment model, otherwise the basic model is refitted.
[0014] As a further optimization of the above-mentioned robust SLAM method based on improved visual feature extraction and matching optimization: The preset set constraints include point feature constraints, line feature constraints and surface feature constraints; The method of optimizing the dynamic environment model based on the preset set constraints using the comprehensive feature fusion module is as follows: Among them, N point , N line , N plane are the number of point features, line features and surface features respectively, and e point , e line , e plane They represent point feature error, line feature error and surface feature error respectively, and minimize the total error E total Optimize the dynamic environment model.
[0015] A highly robust SLAM system based on improved visual feature extraction and matching optimization is used for the above-mentioned highly robust SLAM method based on improved visual feature extraction and matching optimization, and the system comprises: A data acquisition device, used for acquiring the environment image; A data processing device, used to process the environment image using the SLAM model to generate the dynamic environment model; The model output device is used to output the dynamic environment model.
[0016] Beneficial effects: Based on the SuperPoint algorithm, the present invention constructs a SLAM model based on a multi-scale feature fusion module, an attention module, a triplet loss module, a random consensus algorithm module and a comprehensive feature fusion module. In the multi-scale feature fusion, a multi-level feature map is constructed by the feature pyramid network (FPN) of the multi-scale feature fusion module, and features of different depths are integrated by combining upsampling operations and jump connections to ensure the effective fusion of high- and low-resolution features; adaptive weighting is performed by the attention module to enhance the model's attention to key areas, thereby highlighting important features and suppressing irrelevant information; the distance between features of different classes is maximized and the distance between features of the same class is minimized by the triplet loss, thereby improving the distinguishing ability of feature descriptors; the random consensus algorithm (RANSAC) module is integrated to give priority to high-confidence matching pairs to reduce the number of iterations and further improve the accuracy of matching; in the process of fusing point, line and surface features, the geometric constraint optimization model is used to effectively enhance the system's positioning ability in environments where point features are scarce or textures are weak;
[0017] The present invention can handle the fusion and optimization of multi-feature and multi-source data in complex environments, is suitable for complex indoor and outdoor scenes, significantly improves the accuracy, robustness and adaptability of visual inertial SLAM, reduces the impact of external interference on the system, and improves the performance of the system in dynamic and unstructured environments. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 It is a structural schematic diagram of the SLAM model in the present invention; Figure 2 It is a flowchart of the random consistency algorithm module. DETAILED DESCRIPTION
[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0020] like Figure 1 and 2 As shown, the present invention first provides a strong robust SLAM method based on improved visual feature extraction and matching optimization, including S1 to S9.
[0021] S1. Collect environmental images in real time and input the environmental images into a pre-trained SLAM (Simultaneous Localization and Mapping) model. The SLAM model is built based on the SuperPoint algorithm. The SLAM model includes a multi-scale feature fusion module, an attention module, a triplet loss module, a random consensus algorithm module and a comprehensive feature fusion module.
[0022] S2. Use the feature encoder of the SuperPoint algorithm to extract features from the environment image and obtain multiple original feature maps of different depths.
[0023] S3. Using a multi-scale feature fusion module to process the original feature map to generate multiple hierarchical feature maps of different depths, and establishing a feature connection relationship between the hierarchical feature maps to obtain a fused feature. The specific method of using a multi-scale feature fusion module to process the original feature map to generate multiple hierarchical feature maps of different depths and establishing a feature connection relationship between the hierarchical feature maps includes S31 to S33.
[0024] S31. Use a feature pyramid network to construct a feature map pyramid based on the original feature map, where the feature map pyramid includes multiple hierarchical feature maps.
[0025] S32, upsample the hierarchical feature map and connect it horizontally with the original feature map: P i =f(C i )+UpSample(P i+1 ); Among them, P i is the hierarchical feature map of level i, C i is with P i The corresponding original feature map, UpSample(P i+1 ) is upsampling, and f(·) is the convolution operation.
[0026] S33. Constructing fusion features
[0027] S4. Adaptively weighting the hierarchical feature map using the attention module to obtain an enhanced feature map. The method of adaptively weighting the hierarchical feature map using the attention module includes S413.
[0028] S41. Perform global average pooling on each channel in the hierarchical feature map to obtain a global description of each channel: Among them, F∈R H×W×C is the input hierarchical feature map, H, W, C represent the height, width and number of channels of the feature map respectively, z c Represents the global description of the c-th channel.
[0029] S42. The result of global pooling is passed through two fully connected layers in sequence. The first fully connected layer is used to reduce the number of channels of the hierarchical feature map, and the second fully connected layer is used to increase the number of channels after the dimension reduction: s = σ(W2δ(W1z)); Where z∈R C is the global information of all channels, and is the weight matrix of the fully connected layer, δ(·) is the ReLU activation function, σ(·) is the Sigmoid activation function, ensuring that the output weight value is in the range of (0, 1), s∈R C is the adaptive weight of each channel.
[0030] S43, adaptively weighting the hierarchical feature map: where s c is the adaptive weight of the cth channel, To enhance the feature map.
[0031] S5. The feature point decoder of the SuperPoint algorithm is used to detect feature points from the enhanced feature map, and the descriptor decoder of the SuperPoint algorithm is used to generate feature descriptors corresponding to the feature points. There are three types of feature descriptors, namely Anchor, Positive and Negative. After the feature descriptors are generated, the feature descriptors are converted into feature vectors, namely f A ,f P ,f N .
[0032] S6. Optimize feature points and feature descriptors based on distance using a triplet loss module. The triplet loss function of the triplet loss module is: L(A,P,N)=max{0,d(A,P)-d(A,N)+α}; d(A,P) is the distance between Anchor and Positive, d(A,N) is the distance between Anchor and Negative, and α is the predefined safety margin.
[0033] The method for optimizing the descriptor includes S61 to S62.
[0034] S61, with f A Based on the distance metric function, the similarity between feature descriptors is calculated: Among them, P·P2 is the Euclidean norm.
[0035] S62. Obtain the average value of all triple combinations by minimizing the triple loss function: Where i represents the triple combination that traverses all training samples.
[0036] Furthermore, the method to minimize the triple loss function is the gradient descent method, specifically: in, and They are the gradients of Anchor, Positive and Negative respectively.
[0037] S7. Reorder all feature points according to the confidence of the optimized feature points using the random consensus algorithm module, and select multiple feature points as backup data points.
[0038] The method of reordering all feature points according to the confidence of the optimized feature points by using the random consensus algorithm module includes S71 to S73.
[0039] S71. Calculate the confidence of all feature points.
[0040] S72. Sort all feature points in order from high to low confidence levels, and form an ordered set of feature points: D sorted ={d1,d2,...,d n}, so that p1≥p2≥…≥p n ; Among them, p i is the confidence value of the i-th feature point, and p i The value range is [0, 1].
[0041] S73. Select the first m feature points from the ordered set of data points as spare data points.
[0042] S8. Generate a dynamic environment model based on all the spare data points. The method of generating a dynamic environment model based on all the spare data points includes S81 to S83.
[0043] S81. Fit the basic model using all candidate data points. The fitting method is: Among them, θ is the model parameter, x i ,y i are the input and output numbers respectively, f(x i ; θ) represents the use of model parameters θ for input x i Function for making predictions, argmin θ It means finding the parameter set θ that minimizes the error.
[0044] S82, calculate the error between all feature points and the basic model, and screen the feature points based on the error to screen out the inner points relative to the basic model. The error calculation method is Among them, e i is the error of the i-th feature point, indicating the deviation between the data point and the fitted model, is the error function, used to calculate the predicted value of the model and the true value y i The distance between them is usually the Euclidean distance. The model for the data point x i The prediction of i is the true value, and d(·,·) represents the error metric function.
[0045] When the error is less than the preset inlier threshold ∈, the feature point is considered to be an inlier, i.e.
[0046] S83. If the number of inliers selected reaches a preset threshold, the basic model is used as the dynamic environment model. Otherwise, the basic model is refitted and the number of iterations is increased by one. The total number of inliers is expressed as And compare the total number of inliers with the preset threshold. If N inliers >N target , that is, the total number of internal points reaches a preset threshold, then the dynamic environment model can be determined.
[0047] S9. Utilizing the comprehensive feature fusion module to optimize the dynamic environment model based on preset set constraints. The preset set constraints include point feature constraints, line feature constraints, and surface feature constraints.
[0048] Point features are the basic geometric constraints in visual inertial navigation. By identifying and tracking key point features in the image, these two-dimensional points can be mapped to three-dimensional space to perform pose estimation. Point feature constraints are usually reflected by reprojection error, which measures the deviation between the feature point position calculated by the camera pose and the actual observed position. By minimizing the reprojection error, the camera pose can be continuously optimized so that the estimated three-dimensional points and the actual observed points are as consistent as possible, thereby improving the positioning accuracy of the system. Specifically, the point feature constraint is optimized based on the reprojection error, and its calculation method is e point, =Pp-π(P,T)P2, where p is the point feature coordinate in the image, P is the point coordinate in three-dimensional space, π(P,T) is the position of the three-dimensional point P projected onto the image plane through the camera projection model, and T is the camera pose matrix.
[0049] Line features are an important supplement to feature constraints in three-dimensional environments. Especially when there are significant edges in the scene, line features can provide stronger geometric information than point features. These line features can effectively reduce feature matching failures caused by texture loss, thereby providing robust geometric constraints in the scene. Since line features usually span multiple pixels, they are more robust to noise than point features. Therefore, in complex environments, line features can provide more stable geometric information. The core of line feature constraints is to minimize the error after the three-dimensional line segment is projected onto the image plane. The calculation method is: Among them, 1 i represents the coordinates of the i-th line segment in the image, L i is the corresponding line segment in three-dimensional space, π(L i ,T) is the three-dimensional line segment L i The position projected onto the image plane by the camera projection model, dist(.) represents the distance error between line segments, such as the minimum distance from the endpoint to the line.
[0050] The surface features are obtained by plane fitting or multi-view geometry inference to provide geometric constraints between the surface and the camera. The surface feature constraints are usually defined using the plane normal vector and distance error, and the calculation method is e plane = dist1(π(P plane ,T),n), where P plane represents a three-dimensional point on a plane, π(P plane ,T) represents the position of the three-dimensional point projected onto the image plane, n is the normal vector of the plane, indicating the direction of the plane, and dist1(.) represents the distance from the point to the plane.
[0051] The method of optimizing the dynamic environment model based on the preset set constraints using the comprehensive feature fusion module is as follows: Among them, N point , N line , N plane are the number of point features, line features and surface features respectively, and e point , e line , e plane They represent point feature error, line feature error and surface feature error respectively, and minimize the total error E total Optimize the dynamic environment model.
[0052] The present invention further provides a highly robust SLAM system based on improved visual feature extraction and matching optimization, which is used for the above-mentioned highly robust SLAM method based on improved visual feature extraction and matching optimization. The system includes a data acquisition device, a data processing device and a model output device.
[0053] The data acquisition device is used to acquire the environmental image, and the data acquisition device may include multiple cameras with different angles.
[0054] The data processing device is used to process the environment image using the SLAM model to generate the dynamic environment model. The data processing device may include a processor, a memory and a data interface, which are all mature existing technologies and will not be described in detail here.
[0055] The model output device is used to output the dynamic environment model, and the model output device may include a display module.
[0056] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A strong robust SLAM method based on improved visual feature extraction and matching optimization, characterized in that: The method comprises the following steps: collecting environmental images in real time, and inputting the environmental images into a pre-trained SLAM (Simultaneous Localization and Mapping) model, wherein the SLAM model is constructed based on the SuperPoint algorithm, and the SLAM model comprises a multi-scale feature fusion module, an attention module, a triplet loss module, a random consensus algorithm module, and a comprehensive feature fusion module; The feature encoder of the SuperPoint algorithm is used to extract features from the environment image to obtain multiple original feature maps of different depths; A multi-scale feature fusion module is used to generate multiple hierarchical feature maps of different depths from the original feature map, and a feature connection relationship is established between the hierarchical feature maps to obtain fused features; The attention module is used to adaptively weight the hierarchical feature map to obtain an enhanced feature map; The feature point decoder of the SuperPoint algorithm is used to detect feature points from the enhanced feature map, and the descriptor decoder of the SuperPoint algorithm is used to generate feature descriptors corresponding to the feature points; The triplet loss module is used to optimize feature points and feature descriptors based on distance; The random consensus algorithm module is used to reorder all feature points according to the confidence of the optimized feature points, and multiple feature points are selected as backup data points; Generate a dynamic environment model based on all the alternate data points; The dynamic environment model is optimized based on the preset set constraints using the comprehensive feature fusion module.
2. The highly robust SLAM method based on improved visual feature extraction and matching optimization as claimed in claim 1, characterized in that, The specific method of using the multi-scale feature fusion module to generate multiple hierarchical feature maps of different depths from the original feature map and establish a feature connection relationship between the hierarchical feature maps includes: A feature pyramid network is used to construct a feature map pyramid based on the original feature map. The feature map pyramid includes multiple hierarchical feature maps. The hierarchical feature maps are upsampled and laterally connected to the original feature maps: P i =f(C i )+UpSample(P i+1 ); Among them, P i is the hierarchical feature map of level i, C i is with P i The corresponding original feature map, UpSample(P i+1 ) is upsampling, and f(·) is the convolution operation; Constructing fusion features 3. The strong robustness SLAM method based on improved visual feature extraction and matching optimization as claimed in claim 1, characterized in that, Methods for adaptively weighting hierarchical feature maps using attention modules include: Perform global average pooling on each channel in the hierarchical feature map to obtain a global description of each channel: Among them, F∈R H×W×C is the input hierarchical feature map, H, W, C represent the height, width and number of channels of the feature map respectively, z c Represents the global description of the c-th channel; The result of global pooling is passed through two fully connected layers in sequence. The first fully connected layer is used to reduce the number of channels of the hierarchical feature map, and the second fully connected layer is used to increase the number of channels after the reduction: s = σ(W2δ(W1z)); Where z∈R C is the global information of all channels, and is the weight matrix of the fully connected layer, δ(·) is the ReLU activation function, σ(·) is the Sigmoid activation function, ensuring that the output weight value is in the range of (0, 1), s∈R C is the adaptive weight of each channel; Adaptively weight hierarchical feature maps: where s c is the adaptive weight of the cth channel, To enhance the feature map.
4. The strong robustness SLAM method based on improved visual feature extraction and matching optimization as claimed in claim 1, characterized in that, There are three types of feature descriptors, namely Anchor, Positive and Negative. After generating the feature descriptor, the feature descriptor is converted into a feature vector, which are f A ,f P ,f N .
5. The strong robustness SLAM method based on improved visual feature extraction and matching optimization as claimed in claim 4, characterized in that, The triplet loss function of the triplet loss module is: L(A,P,N)=max{0,d(A,P)-d(A,N)+α}; d(A,P) is the distance between Anchor and Positive, d(A,N) is the distance between Anchor and Negative, and α is the predefined safety margin; With f A Based on the distance metric function, the similarity between feature descriptors is calculated: Among them, P·P2 is the Euclidean norm; The average value of all triple combinations is obtained by minimizing the triple loss function: Where i represents the triple combination that traverses all training samples.
6. The strong robustness SLAM method based on improved visual feature extraction and matching optimization as claimed in claim 5, characterized in that, The method to minimize the triple loss function is the gradient descent method, specifically: in, and They are the gradients of Anchor, Positive and Negative respectively.
7. The robust SLAM method based on improved visual feature extraction and matching optimization as claimed in claim 1, characterized in that, The method of reordering all feature points according to the confidence of the optimized feature points using the random consensus algorithm module includes: Calculate the confidence of all feature points; Sort all feature points in order from high to low confidence and form an ordered set of feature points: D sorted ={d1,d2,...,d n }, so that p1≥p2≥…≥p n ; Among them, p i is the confidence value of the i-th feature point, and p i The value range of is [0,1]; The first m feature points are selected from the ordered set of data points as backup data points.
8. The robust SLAM method based on improved visual feature extraction and matching optimization as claimed in claim 1, characterized in that, Methods for generating a dynamic environment model based on all the alternate data points include: All candidate data points are used to fit the basic model. The fitting method is: Among them, θ is the model parameter, x i ,y i are the input and output numbers respectively, f(x i ; θ) represents the use of model parameters θ for input x i Function for making predictions, argmin θ It means finding the parameter set θ that minimizes the error; Calculate the error between all feature points and the basic model, and screen the feature points based on the error to select the internal points relative to the basic model; If the number of filtered inliers reaches a preset threshold, the basic model is used as the dynamic environment model, otherwise the basic model is refitted.
9. The strong robustness SLAM method based on improved visual feature extraction and matching optimization as claimed in claim 1, characterized in that, The preset set constraints include point feature constraints, line feature constraints and surface feature constraints; The method of optimizing the dynamic environment model based on the preset set constraints using the comprehensive feature fusion module is as follows: Among them, N point , N line , N plane are the number of point features, line features and surface features respectively, and e point , e line , e plane They represent point feature error, line feature error and surface feature error respectively, and minimize the total error E total Optimize the dynamic environment model.
10. A highly robust SLAM system based on improved visual feature extraction and matching optimization, characterized in that: For implementing the highly robust SLAM method based on improved visual feature extraction and matching optimization as described in any one of claims 1 to 9, the system comprises: A data acquisition device, used for acquiring the environment image; A data processing device, used to process the environment image using the SLAM model to generate the dynamic environment model; The model output device is used to output the dynamic environment model.